Budget shell calls like CLI, and when rounds are exhausted inject a no-more-tools system nudge plus any user_facing_hint constraints into the final synthesis pass.
Co-authored-by: Cursor <cursoragent@cursor.com>
Wire should_stop from aborted runs, snapshot abort before clearing flags so finals are not emitted after stop, and keep fail-path turn_uuid aligned for bubble splitting.
Co-authored-by: Cursor <cursoragent@cursor.com>
Reuse channel_session_v2 per job_id so cron runs do not create a new
oclaw session each time; isolate scheduled turn context to the active
turn_uuid so shared sessions do not leak prior run history.
Co-authored-by: Cursor <cursoragent@cursor.com>
Pass the active workspace lane into tool execution and default schedule_create to that specialist so WhatsApp ops-bound jobs are stored and run as ops instead of generalist.
Co-authored-by: Cursor <cursoragent@cursor.com>
Wire mention JIDs/names through scheduled delivery and immediate replies (including attachment captions), sanitize group message text for the model, and surface channel-derived sessions in admin Chat.
Co-authored-by: Cursor <cursoragent@cursor.com>
Introduce persisted cron/interval jobs, gateway scheduler loop, worker turns,
and admin CRUD/edit UI. Route proactive reminders via the originating chat
channel (WhatsApp vs WeChat), harden Weixin outbound with durable queue and
PG-compatible polling, and fix chat UI to show each scheduled reminder separately.
Co-authored-by: Cursor <cursoragent@cursor.com>
Raise Admin/UI defaults and backend clamps for AIA_TURN_MAX_TOOL_ROUNDS across gateway, worker, and direct loop paths.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add last-chance DSML promote on combined content+reasoning and parse again before protocol_mismatch so split-field DeepSeek output runs real tools and produces tool_result rows.
Co-authored-by: Cursor <cursoragent@cursor.com>
Parse DSML from combined content+reasoning when markers span fields; show a clear mismatch message when parse fails; run finalize synthesis when tools executed but the assistant body is empty.
Co-authored-by: Cursor <cursoragent@cursor.com>
Parse <｜｜DSML｜｜> markup, add DSML recovery to openai_responses transport, and avoid persisting raw DSML as plain assistant text when promotion fails.
Co-authored-by: Cursor <cursoragent@cursor.com>
Promote DSML markup from assistant text into native tool calls during streaming and completion, filter DSML from UI tokens, and document AIA_DSML_TEXT_TOOLS for local vLLM proxies.
Co-authored-by: Cursor <cursoragent@cursor.com>
Append english_output_guard when lang=en, strengthen ops/generalist EN role prompts,
pass NETX_TOOL_LANG into netx tools, and map Chinese protocol buckets to English.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Rename platform/ to svc/ to avoid shadowing stdlib platform.
- Replace from oclaw.* with from svc/runtime/interfaces; update -m CLI paths.
- tests/conftest: prepend repo root to sys.path (no parent-folder package name).
- CI: paths and offline_eval script under repo root.
- Ops scripts: PYTHONPATH must be repo root for python -m runtime.* (fixes gateway/WhatsApp sidecar startup).
- Fix default oclaw.json path in tabular/file attachment limits; stabilize attachment test config.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add dsml_tool_parse per upstream HF encoding README (invoke/parameter,
string true|false, JSON for non-strings). Normalize <||DSML|| gateways.
Wire into run_oclaw_direct_loop when AIA_DSML_TEXT_TOOLS is on or
base_url/model suggests DeepSeek; skip DSML repair retry in that mode.
Strip the first tool_calls block from persisted assistant body. Finalize
pass keeps DSML disabled so no tools run after the no-tool round.
Tests cover parser variants and direct_loop execution for DeepSeek URL
and forced env.
Co-authored-by: Cursor <cursoragent@cursor.com>
Non-thinking profiles previously wrote model reasoning only as separate
event_type=reasoning rows whose event_payload held chunk_index metadata, so
JSON inspection showed tool payloads but no reasoning_content key.
Store the full merged reasoning blob on the primary assistant row whenever
present (same shape as thinking mode). Update instruction-leak filtering to
scan event_payload.reasoning_content as well as legacy reasoning rows.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Forward reasoning_content and non-stream reasoning to on_token in
OpenAI chat completions; Gemini thought parts to on_token.
- Stream bubble: always render session.tool panels; format tool_use_call
payloads and title as call · tool_name.
- Emit tool_use_call via on_tool_ui before tool execution so WS shows
invoke cards before tool_use_result.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Resolve legacy multimodal inputs from session history when the user sends
text-only (assistant images first, then user uploads); optional env toggles.
- Support relay_pointer image resolution in legacy lane collector.
- Align uvicorn WebSocket frame limit with MAX_PAYLOAD_BYTES; document
OCLAW_UVICORN_WS_MAX_SIZE and compatible-mode env knobs.
- Admin chat: encode attachments after bubble preview; FileReader-based
base64 for large files.
- Tests and docs for new behavior.
Co-authored-by: Cursor <cursoragent@cursor.com>
- Add runtime/plan_agent_v2 package and shims; gateway/direct_loop/WS wiring
- Admin chat: interaction mode and specialist only in user menu; session API stores memory_mode and execution_mode only
- POST /admin/api/chat/user-mode mirrors plan_agent_version to AIA_EXPERT_PLAN_AGENT_V2_ENABLED (v2 to 1, v1 to 0)
- Composer cleanup (hidden mode select, no reasoning toggle in meta bar); tests and _local/system.env.example
Co-authored-by: Cursor <cursoragent@cursor.com>
This preserves multimodal fidelity for the latest user turn while preventing historical/tool replay bloat, and adds admin UI support for referenced attachment preview/download flows.
Made-with: Cursor
